Offline Reinforcement Learning for Learning to Dispatch for Job Shop Scheduling at ECML PKDD 2025
Date:
I presented Offline-LD in the Journal Track of ECML PKDD 2025, the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, held in Porto from 15–19 September 2025. The paper appeared in Machine Learning (Springer) and was selected for presentation at the conference through the journal track.
The talk covered Offline-LD, the first fully end-to-end offline RL approach for the Job Shop Scheduling Problem: how maskable variants of Quantile Regression DQN and discrete Soft Actor–Critic can be trained through Conservative Q-Learning to learn dispatching policies from historical schedules alone, with no simulator, and why the resulting policies outperform online RL baselines while training on only 100 constraint-programming solutions.
The finding that drew most discussion was that deliberately adding noise to the expert dataset did not degrade performance, and in places improved it — which suggests these methods are more robust to the imperfections of real industrial scheduling logs than the offline RL literature’s emphasis on data quality would predict.
Links